Integrating Domain Knowledge into AI‑Driven Medical Reasoning: Trends, Challenges, and Techniques

Guest speaker: Su Jia, PhD, Huawei Cloud. Editor: Xiong Danni, Wuhan Tianyu Information. Platform: DataFunTalk.

Guide: With the popularity of artificial intelligence, especially deep learning, the medical industry is experiencing a revolutionary wave. This article discusses how to embed domain knowledge into medical reasoning tasks, providing an effective AI + medical solution. Main contents include: smart healthcare status, core issues, domain knowledge, and knowledge‑enhanced reasoning.

01. Smart Healthcare Status

1. Prospects

Statistical data shows that China’s medical software market has been growing at about 50% annually. Forecasts predict the global smart‑healthcare market will reach $27.5 billion by 2025, with China and the United States leading development.

China faces uneven medical resource distribution, creating an urgent need for AI assistance. Current AI is transitioning from perception to cognition, spawning many image‑based AI applications such as cancer screening. The COVID‑19 pandemic has accelerated industry growth, prompting supportive policies and strong involvement from major vendors, including Huawei Cloud’s natural advantage in cloud computing.

2. Application Scenarios

Established smart‑healthcare projects include medical image diagnosis, disease screening and prediction, clinical decision support, and intelligent medical records. These illustrate the hot market for AI + medical. Core AI technologies involved are computer vision, natural language processing, speech recognition, and machine learning. Deep learning has increased the frequency of neural‑network algorithms in medical literature.

3. Important Events

The timeline shows key milestones: 1960s knowledge‑base attempts to simulate doctor reasoning; 1978 Beijing Traditional Chinese Medicine Hospital’s “Guan Youbo Liver Disease” program; 2000 Da Vinci surgical robot approval; 2011 IBM Watson; recent rapid emergence of AI + medical products.

02. Core Issues

The industry faces three main challenges:

Industry constraints: medical data is highly sensitive, leading to cautious development and a “Matthew effect” where large hospitals attract most resources, leaving remote areas underserved. AI must help balance this disparity while respecting stakeholder interests.

Data: Compared with developed countries, China lags in health‑data management. Regulations require signed usage agreements, ethics committee approval, and research‑only usage. Additional challenges include privacy protection, data quality, storage, and security.

Algorithms & compute: AI models are often black‑boxes with limited interpretability. Clinicians demand evidence chains for predictions. Sparse data, high computational costs of large models, and rapid hardware evolution further complicate deployment.

03. Domain Knowledge

1. Medical Knowledge Graphs

Construction follows standard graph‑building steps: design of a medical ontology, text acquisition, entity and attribute extraction, relation extraction, and knowledge alignment.

Extraction methods have evolved through four stages: rule‑based/dictionary, traditional machine learning (e.g., CRF, SVM), deep learning, and pre‑training models.

Relation extraction work includes adding positional information to CNN encoders, integrating syntactic trees into BiLSTM‑RNN frameworks, and reformulating relation extraction as an entity‑recognition task.

Knowledge alignment now often leverages pre‑trained models to compute similarity between entity embeddings and dictionary entries, as demonstrated by research from Cambridge and Amazon.

Prominent medical knowledge graphs include CMeKG, OMAHA, and recent COVID‑19 graphs from Huawei Cloud and Zhejiang University. The I2B2 benchmark is a widely used English evaluation dataset.

2. Medical Pre‑training Models

Since BERT’s 2018 debut, domain‑specific pre‑training has flourished. BioBERT, trained on biomedical text, outperforms generic models on entity recognition, relation extraction, and QA tasks.

MC‑BERT (Alibaba) uses full‑masking of medical entities and phrases; MT‑BERT applies multi‑task learning across downstream tasks; BERT‑MK (Huawei Noah Lab) injects knowledge‑graph semantics via entity‑mediated representations.

Huawei Cloud and Noah also propose a model that jointly learns entity‑mask and relation‑classification tasks, integrating both entity and relation knowledge into the pre‑training process.

04. Knowledge‑Enhanced Reasoning Techniques

1. Medical Knowledge Computation

AI reasoning comprises two aspects:

Knowledge reasoning (generating new knowledge) : Requires a complete evidence chain and interpretability. Traditional knowledge‑completion methods (TransE, RotatE) have been used to enrich COVID‑19 graphs for treatment discovery.

Model prediction after training : Handles tasks such as QA, literature retrieval, and recommendation. Sparse medical graphs are completed using pre‑trained models that convert relational inference into numerical computation.

Examples include re‑training BERT on disease‑related Wikipedia sentences to boost QA performance, and Amazon’s COVID‑19 literature graph for similarity‑based retrieval.

Classification tasks (e.g., disease diagnosis) benefit from knowledge graphs: Zhengzhou University & Pengcheng Lab combine obstetric graphs with entity alignment for diagnosis; Texas‑A&M & Pengcheng Lab’s Med‑BERT improves heart‑failure and pancreatic‑cancer complication prediction.

Drug recommendation scenarios also leverage GNN + pre‑training models to predict medication usage.

Huawei proposes a framework that encodes patient text with a pre‑trained model, extracts entities via a knowledge graph, and jointly models context and entity representations for downstream clinical tasks.

05. Q&A

Q: How is Domain‑Bert trained? Any architectural changes?

A: It follows a RoBERTa‑style framework; the architecture remains largely unchanged, with adaptations for Chinese medical data.

Q: Do age and allergy information affect medication prediction?

A: Experiments show age and gender improve prediction by about 3% on our dataset.

Q: What recall rate is considered good for pre‑training? How to decide sample size?

A: Current models achieve 70‑80% recall; the target depends on the task and data quality. Sample size is usually guided by prior work, often scaling an order of magnitude larger.

Q: How many GPUs are used for pre‑training?

A: Eight V100 GPUs (32 GB each).

Q: Is data cleaning performed with a pre‑trained model?

A: No. Data is cleaned first using templates, rules, and machine‑learning methods (e.g., MC‑BERT for weak‑label quality classification) before training.

Thank you for reading.

At the end of the article, please like, share, and give a "3‑click" support.

Share guest: Su Jia.

Event recommendation: On December 18, DataFunCon 2021 year‑end conference will feature Meituan senior algorithm expert Wang Sirui presenting a knowledge‑graph forum with guests from Peking University, Southeast University, Baidu, Xiaomi, and Meituan. Scan the QR code to register for free.